Traditional Chinese Text Recognition Dataset: Synthetic Dataset and Labeled Data

Authors: Yi-Chang Chen, Yu-Chuan Chang, Yen-Cheng Chang and Yi-Ren Yeh

Paper: https://arxiv.org/abs/2111.13327

Scene text recognition (STR) has been widely studied in academia and industry. Training a text recognition model often requires a large amount of labeled data, but data labeling can be difficult, expensive, or time-consuming, especially for Traditional Chinese text recognition. To the best of our knowledge, public datasets for Traditional Chinese text recognition are lacking.

We generated over 20 million synthetic data and collected over 7,000 manually labeled data TC-STR 7k-word as the benchmark. Experimental results show that a text recognition model can achieve much better accuracy either by training from scratch with our generated synthetic data or by further fine-tuning with TC-STR 7k-word.

Synthetic Dataset: TCSynth

Inspired by MJSynth, SynthText and Belval/TextRecognitionDataGenerator, we propose a framework for generating scene text images for Traditional Chinese. To produce synthetic text images similar to real-world ones, we use different kinds of mechanisms for rendering, including word sampling, character spacing, font types/sizes, text coloring, text stroking, text skewing/distorting, background rendering, text Location and noise.

TCSynth dataset includes 21,535,590 synthetic text images.

• Download:
  • TCSynth LMDB Format (Recommanded) [21 GB (54 GB after untaring)]
  • TCSynth Raw Format [18 GB (85 GB after untaring)]

TCSynth-VAL dataset includes 6,000 synthetic text images for validation.

• Download:
  • TCSynth-VAL LMDB Format (Recommanded)
  • TCSynth-VAL Raw Format

LMDB Format

After untaring,

TCSynth/
├── data.mdb
└── lock.mdb

Our data structure of LMDB follows the repo. clovaai/deep-text-recognition-benchmark. The value queried by key 'num-samples'.encode() gets total number of text images. The indexes of text images starts from 1. Given the index, we can query binary of the image and its label by key 'image-%09d'.encode() % index and 'label-%09d'.encode() % index. The implement details are shown in the class LmdbConnector in lmdb_tools/lmdb_connector.py.

We also provide several tools to manipulate the LMDB shown in lmdb_tools. Before using those tools, we should install some dependencies. (tested with python 3.6)

pip install -r lmdb_tools/requirements.txt

• Insert images into LMDB

python lmdb_tools/prepare_lmdb.py \
  --input_dir IMG_FOLDER \
  --gt_file GT \
  --output_dir LMDB_FOLDER

• Insert images into LMDB (asynchronous version)

python lmdb_tools/prepare_lmdb_async.py \
  --input_dir IMG_FOLDER \
  --gt_file GT \
  --output_dir LMDB_FOLDER \
  --workers WORKERS

• Extract images from LMDB (asynchronous version) (convert LMDB Format to Raw Format)

python lmdb_tools/extract_to_files.py \
  --input_lmdb LMDB_FOLDER \
  --output_dir IMG_FOLDER \
  --workers WORKERS

Raw Format

After untaring,

TCSynth_raw/
├── labels.txt
├── 0000/
│   ├── 00000001.jpg
│   ├── 00000002.jpg
│   ├── 00000003.jpg
│   └── ...
├── 0001/
├── 0002/
└── ...

format of labels.txt: {imagepath}\t{label}\n, for example:

0000/00000001.jpg 㒓
...

Labeled Data: TC-STR 7k-word

Our TC-STR 7k-word dataset collects about 1,554 images from Google image search to produce 7,543 cropped text images. To increase the diversity in our collected scene text images, we search for images under different scenarios and query keywords. Since the collected scene text images are to be used in evaluating text recognition performance, we manually crop text from the collected images and assign a label to each cropped text box.

TC-STR 7k-word dataset includes a training set of 3,837 text images and a testing set of 3,706 images.

• Download: TC-STR.tar.gz

After untaring,

TC-STR/
├── train_labels.txt
├── test_labels.txt
└── images/
    ├── xxx_1.jpg
    ├── xxx_2.jpg
    ├── xxx_3.jpg
    └── ...

format of xxx_labels.txt: {imagepath}\t{label}\n, for example:

images/billboard_00000_010_雜貨鋪.jpg 雜貨鋪
images/sign_02616_999_民生路.png 民生路
...

Citation

Please consider citing this work in your publications if it helps your research.

@article{chen2021traditional,
  title={Traditional Chinese Synthetic Datasets Verified with Labeled Data for Scene Text Recognition},
  author={Yi-Chang Chen and Yu-Chuan Chang and Yen-Cheng Chang and Yi-Ren Yeh},
  journal={arXiv preprint arXiv:2111.13327},
  year={2021}
}